The recognition effect based on text_prompt is poor

If you enter the keyword 'beverage', it can be recognized normally . However, when it comes to prompts like 'The white beverage on the upper shelf' or 'White and transparent beverage', the recognition performance is not very good .

I further discovered that the model made incorrect judgments regarding adjectives such as "larger", "smaller", "left", "right", or "first layer" and "second layer". For example, it misidentified the yellow beverage on the left of the blue beverage as belonging to the first column.

Then I tried to label the drinks with numbers. When the prompt word is 'number', it can recognize all numbers normally, but it fails to recognize specific numbers like 123456

The issue names no files, tests, or entry points. Start by reproducing the reported behavior with the attached beverage images and prompts such as “beverage,” “The white beverage on the upper shelf,” positional adjectives, and specific numbers; trace the relevant inference path to determine whether the behavior is actionable. Done requires a confirmed cause and a validated correction or clearly documented model limitation.

Written by the indexing model from the issue text.